Source-linked AI summary
DebtRank: A microscopic foundation for shock propagation
Marco Bardoscia, Stefano Battiston, Fabio Caccioli, Guido Caldarelli
TL;DR
The paper develops a microscopic theory of financial-network instability to address DebtRank’s potential underestimation of losses. It derives shock dynamics from balance-sheet identities, links stability to the interbank leverage matrix, and finds substantial shock amplification in European-bank data.
Problem
DebtRank can underestimate systemic losses because banks propagate distress only once, especially in networks containing loops.
Method
The paper iterates individual banks’ balance-sheet identities and assumes that borrowers’ relative equity changes produce equal relative changes in lenders’ interbank assets.
Results
Network effects amplify exogenous shocks by factors ranging from three in normal periods to six during the crisis under a 0.5% external-asset shock to all banks.
Takeaways & Limitations
Stability depends on the interbank leverage matrix: shocks are damped when its largest eigenvalue modulus is below one and amplified until default when it exceeds one.
Takeaways & Limitations
Allowing the original DebtRank dynamics to propagate shocks while equity remains positive would double-count losses.
Abstract
from arXiv · showhide
The DebtRank algorithm has been increasingly investigated as a method to estimate the impact of shocks in financial networks, as it overcomes the limitations of the traditional default-cascade approaches. Here we formulate a dynamical "microscopic" theory of instability for financial networks by iterating balance sheet identities of individual banks and by assuming a simple rule for the transfer of shocks from borrowers to lenders. By doing so, we generalise the DebtRank formulation, both providing an interpretation of the effective dynamics in terms of basic accounting principles and preventing the underestimation of losses on certain network topologies. Depending on the structure of the interbank leverage matrix the dynamics is either stable, in which case the asymptotic state can be computed analytically, or unstable, meaning that at least one bank will default. We apply this framework to a dataset of the top listed European banks in the period 2008 - 2013. We find that network effects can generate an amplification of exogenous shocks of a factor ranging between three (in normal periods) and six (during the crisis) when we stress the system with a 0.5% shock on external (i.e. non-interbank) assets for all banks.
Introduction
Financial networks expose interconnectedness and systemic importance, but their structure can also amplify distress. This paper develops a microscopic framework linking bank balance sheets and shock propagation to improve systemic-risk analysis.
- Network representations quantify interconnectedness and identify systemically important institutions through centrality.
- Greater interconnectedness can reduce individual risk under independent shocks while increasing systemic risk through distress amplification.
- Default-based network models usually propagate shocks only after institutions are removed through default events.
- DebtRank addresses this limitation by modelling the incremental accumulation of distress before defaults occur.
- The paper derives DebtRank-like dynamics from individual balance-sheet identities and borrower-to-lender shock transmission, while linking stability to network structure.
Results
The model derives contagion dynamics from bank balance sheets and borrower-to-lender shock transmission, then evaluates systemic losses in European banks from 2008 to 2013. The dynamics can underestimate losses less than original DebtRank and exhibits stability governed by the interbank leverage matrix.
- Model description: The model marks a bank as defaulted when its equity is nonpositive and removes its interbank links from subsequent dynamics.Interbank assets are valued mark-to-market, while liabilities retain face value and defaulted borrowers’ creditors recover nothing.
- Model description: The dynamics generalizes DebtRank by allowing banks to propagate every received shock, preventing underestimation on networks with loops.The two approaches agree on some networks, such as trees, while the proposed dynamics gives larger losses in other cases.
- Model description: Stability depends on the largest eigenvalue of the reduced interbank leverage matrix: values below one damp shocks, whereas values above one amplify them until at least one bank defaults.After defaults modify the matrix, the reduced system may become stable and converge.
- Application to the European banking system: The empirical application reconstructs interbank networks for 183 publicly traded European banks from 2008 to 2013 using balance-sheet totals.The available data provide aggregate interbank borrowing and lending, so individual matrix entries are inferred through a two-step reconstruction.
- Application to the European banking system: A 0.5% external-asset shock produced contagion losses larger than original DebtRank estimates by factors ranging from 1.3 in 2008 to 1.7 in 2013.The comparison covers shocks of 0.5% and 1% applied simultaneously across banks.
- Application to the European banking system: In 2008, relative equity losses reached saturation for external-asset shocks as small as 0.5%, whereas 2013 required shocks five times larger to reach similar losses.The saturation analysis varies the shock from 0.5% to 5.5%.
Discussion
The paper derives a balance-sheet-based generalization of DebtRank that prevents underestimating losses from repeated shock propagation. Applied to 183 European banks, it finds that amplification decreases from 2008 to 2013 and systemic risk is concentrated in banks that are both dangerous and vulnerable.
- Discussion: The generalized dynamics derives from balance-sheet identities, extends DebtRank with further shock propagations, and makes amplification depend on the largest eigenvalue of interbank leverage.Losses attenuate when |λmax| < 1; when |λmax| > 1, a small shock can cause at least one bank to default.
- Discussion: Among 183 European publicly traded banks, shock amplification consistently decreases from 2008 to 2013, with small 2008 shocks sufficiently large to significantly distress all banks.The study also evaluates each bank’s systemic impact and vulnerability through one-at-a-time stress tests.
- Discussion: The most dangerous banks are also the most vulnerable, concentrating systemic risk in a few key players.The paper identifies these banks as targets for effective macroprudential regulation policies.
Methods
The methods represent banks through balance sheets and propagate marked-to-market interbank losses using a reduced leverage matrix. Stability depends on its largest eigenvalue, while the network is reconstructed from balance-sheet data using a fitness model and RAS weighting.
- Methods: A bank’s equity equals assets minus liabilities, and default is represented by equity becoming nonpositive.The model distinguishes interbank from external assets and liabilities, with interbank loans linking lender and borrower balance sheets.
- Methods: The equity-loss dynamics assumes unchanged external items and face-value interbank liabilities while interbank assets are marked to market.The resulting evolution tracks cumulative relative equity loss and removes defaulted banks through the reduced leverage matrix.
- Relation to DebtRank: The generalized dynamics avoids the original DebtRank’s one-time propagation and provides a lower bound comparison because original DebtRank can underestimate cumulative losses.On a unique path through the network, however, each node changes only once, making the result equivalent to original DebtRank.
- Stability properties: The fixed point is stable when |λmax| < 1, whereas |λmax| > 1 causes increasingly large losses and at least one default independently of the initial shock.Between defaults, the same rule applies with the appropriate reduced leverage matrix; subsequent defaults can eventually restore convergence.
- Stability properties: Allowing banks to propagate shocks while their equity remains positive would double-count losses relative to the proposed dynamics.Under the no-default simplification, the alternative iteration produces a quantity always larger than the sum used in the proposed formulation.
- Data: The empirical network uses 183 European publicly traded banks from 2008–2013, with links generated by a directed fitness model at 5% density and weights assigned by RAS.The procedure draws 100 networks and rescales liabilities so total interbank assets equal total interbank liabilities.
Author contributions statement
The authors report distinct roles in developing the conceptual framework, implementing the numerics, interpreting results, and writing the manuscript.
- Author contributions: MB and FC developed the conceptual framework, while MB wrote the code and performed the numerics.MB, SB, FC, and GC contributed to interpreting results and writing the manuscript.
Additional information
The analysis was developed in Python, and the source code is available upon request.
- Additional information: The authors declare no competing financial interests and make the Python source code available upon request.The statement combines the conflict-of-interest declaration with the code-availability information.